Type 1 diabetes mellitus (T1D) is a chronic disease characterized by deficient insulin production and that requires daily administration of this hormone. The cause of T1D is not known, and it is not currently preventable. Even though T1D can be diagnosed at any age, peaks in diagnosis usually occur between the ages of 5 and 7 and at puberty. According to the International Diabetes Federation, T1D affects more than 1.1 million children and adolescents. The study of an extracorporeal system for Blood Glucose Level control, the Artificial Pancreas (AP), began in 1970. The Artificial Pancreas is a closed-loop system composed of a subcutaneous Glucose Continuous Monitoring (CGM) system, a subcutaneous insulin pump and a control algorithm designed to regulate real-time glucose concentration via insulin administration. The control algorithm adapts the therapy on the basis of the real-time state of the patient. The two most critical challenges that the AP deals with are: 1) the intra-patient variability: the BGL fluctuates differently both intra-day and inter-day, depending on heterogeneous factors not directly measurable (e.g., physical activity, lifestyle) or well- known physiological effects, such as the dawn phenomena or daily changes in insulin sensitivity; 2) the inter-patient variability: each patient may have peculiar characteristics depending on its co-morbidities and habits. These variabilities are particularly accentuated in the child and adolescent populations. One of the most promising approaches applied to the glucose control problem is Model Predictive Control (MPC), which is based on glucose predictions obtained through a glucose- insulin model. The quality of these predictions affects the overall performance of the control strategy. MPC needs the knowledge of a mathematical model of the patient to adequately forecast and control her/his blood glucose level. Such models are in general difficult to obtain, and in most of the cases, they are not very precise. The goal of the project is to design learning-based MPC algorithms for the AP, that do not need an explicit model of the patient, but exploits data collected offline and online to understand the dynamics of the process. Different machine learning techniques will be explored, such as Neural Network, Kinky Inference, Gaussian Processes, to reach a patient- tailored control algorithm.